Using Cognates in a French - Romanian Lexical Alignment System: A
Comparative Study
Mirabela Navlea
Linguistique, Langues, Parole (LiLPa)
Université de Strasbourg
22, rue René Descartes
BP, 80010, 67084 Strasbourg cedex
[email protected]
Abstract
This paper describes a hybrid French - Romanian cognate identification module. This module is used by a lexical alignment system. Our
cognate identification method uses lemmatized, tagged and sentence-aligned parallel
corpora. This method combines statistical
techniques, linguistic information (lemmas,
POS tags) and orthographic adjustments. We
evaluate our cognate identification module and
we compare it to other methods using pure statistical techniques. Thus, we study the impact
of the used linguistic information and the orthographic adjustments on the results of the
cognate identification module and on cognate
alignment. Our method obtains the best results
in comparison with the other implemented statistical methods.
1
Introduction
We present a new French - Romanian cognate
identification module, integrated into a lexical
alignment system using French - Romanian parallel law corpora.
We define cognates as translation equivalents
having an identical form or sharing orthographic
or phonetic similarities (common etymology,
borrowings). Cognates are very frequent between
close languages such as French and Romanian,
two Latin languages with a rich morphology. So,
they represent important lexical cues in a French
- Romanian lexical alignment system.
Few linguistic resources and tools for Romanian
(dictionaries, parallel corpora, MT systems) are
currently available. Some lexically aligned corpora or lexical alignment tools (Tufiş et al.,
2005) are available for Romanian - English or
Amalia Todiraşcu
Linguistique, Langues, Parole (LiLPa)
Université de Strasbourg
22, rue René Descartes
BP, 80010, 67084 Strasbourg cedex
[email protected]
Romanian - German (Vertan and Gavrilă, 2010).
Most of the cognate identification modules used
by these systems are purely statistical. As far as
we know, no cognate identification method is
available for French and Romanian.
Cognate identification is a difficult task due to
the high orthographic similarities between bilingual pairs of words having different meanings.
Inkpen et al. (2005) develop classifiers for
French and English cognates based on several
dictionaries and manually built lists of cognates.
Inkpen et al. (2005) distinguish between:
- cognates (liste (FR) - list (EN));
- false friends (blesser (‘to injure’) (FR) - bless
(EN));
- partial cognates (facteur (FR) - factor or mailman (EN));
- genetic cognates (chef (FR) - head (EN));
- unrelated pairs of words (glace (FR) - ice (EN)
and glace (FR) - chair (EN)).
Our cognate detection method identifies cognates, partial and genetic cognates. This method
is used especially to improve a French - Romanian lexical alignment system. So, we aim to obtain a high precision of our cognate identification
method. Thus, we eliminate false friends and
unrelated pairs of words combining statistical
techniques and linguistic information (lemmas,
POS tags). We use a lemmatized, tagged and
sentence-aligned parallel corpus. Unlike Inkpen
et al. (2005), we do not use other external resources (dictionaries, lists of cognates).
To detect cognates from parallel corpora, several
approaches exploit the orthographic similarity
between two words of a bilingual pair. An efficient method is the 4-gram method (Simard et
al., 1992). This method considers two words as
cognates if their length is greater than or equal to
4 and at least their first 4 characters are common.
Other methods exploit Dice’s coefficient (Adam247
Proceedings of Recent Advances in Natural Language Processing, pages 247–253,
Hissar, Bulgaria, 12-14 September 2011.
son and Boreham, 1974) or a variant of this coefficient (Brew and McKelvie, 1996). This measure computes the ratio between the number of
common character bigrams of the two words and
the total number of two word bigrams. Also,
some methods use the Longest Common Subsequence Ratio (LCSR) (Melamed, 1999; Kraif,
1999). LCSR is computed as the ratio between
the length of the longest common substring of
ordered (and not necessarily contiguous) characters and the length of the longest word. Thus,
two words are considered as cognates if LCSR
value is greater than or equal to a given threshold. Similarly, other methods compute the distance between two words, which represents the
minimum number of substitutions, insertions and
deletions used to transform one word into another (Wagner and Fischer, 1974). These methods
use exclusevly statistical techniques and they are
language independent.
On the other hand, other methods use the phonetic distance between two words belonging to a
bilingual pair (Oakes, 2000). Kondrak (2009)
identifies three characteristics of cognates: recurrent sound correspondences, phonetic similarity
and semantic affinity.
Thus, our method exploits orthographic and phonetic similarities between French - Romanian
cognates. We combine n-grams methods with
linguistic information (lemmas, POS tags) and
several input data disambiguation strategies
(computing cognates’ frequencies, iterative extraction of the most reliable cognates and their
deletion from the input data). Our method needs
no external resources (bilingual dictionaries), so
it could easily be extended to other Romance
languages. We aim to obtain a high accuracy of
our method to be integrated in a lexical alignment system. We evaluate our method and we
compare it with pure statistical methods to study
the influence of used linguistic information on
the final results and on cognate alignment.
In the next section, we present the parallel corpora used for our experiments. In section 3, we
present the lexical alignment method. We also
describe our cognate identification module in
section 4. We present the evaluation of our method and a comparison with other methods in
section 5. Our conclusions and further work figure in section 6.
2
The Parallel Corpus
In our experiments, we use a legal parallel corpus (DGT-TM1) based on the Acquis Communautaire corpus. This multilingual corpus is available in 22 official languages of EU member states.
It is composed of laws adopted by EU member
states since 1950. DGT-TM contains 9,953,360
tokens in French and 9,142,291 tokens in Romanian.
We use a test corpus of 1,000 1:1 aligned complete sentences (starting with a capital letter and
finishing with a punctuation sign). The length of
each sentence has at most 80 words. This test
corpus contains 33,036 tokens in French and
28,645 in Romanian.
We use the TTL2 tagger available for Romanian
(Ion, 2007) and for French (Todiraşcu et al.,
2011) (as Web service3). Thus, the parallel corpus is tokenized, lemmatized, tagged and annotated at chunk level.
The tagger uses the set of morpho-syntactic descriptors (MSD) proposed by the Multext
Project4 for French (Ide and Véronis, 1994) and
for Romanian (Tufiş and Barbu, 1997). In the
Figure 1, we present an example of TTL’s output: lemma attribute represents the lemmas of
lexical units, ana attribute provides morphosyntactic information and chunk attribute marks
nominal and prepositional phrases.
<seg lang="FR"><s id="ttlfr.3">
<w lemma="voir" ana="Vmps-s">vu</w>
<w lemma="le" ana="Da-fs"
chunk="Np#1">la</w>
<w lemma="proposition" ana="Ncfs"
chunk="Np#1">proposition</w>
<w lemma="de" ana="Spd"
chunk="Pp#1">de</w>
<w lemma="le" ana="Da-fs"
chunk="Pp#1,Np#2">la</w>
<w lemma="commission" ana="Ncfs"
chunk="Pp#1,Np#2">Commission
</w>
<c>;</c>
</s></seg>
Figure 1 TTL’s output for French (in XCES format)
1
2
http://langtech.jrc.it/DGT-TM.html
Tokenizing, Tagging and Lemmatizing free running texts
3
https://weblicht.sfs.uni-tuebingen.de/
4
http://aune.lpl.univ-aix.FR/projects/multext/
248
3
Lexical Alignment Method
The cognate identification module is integrated
in a French - Romanian lexical alignment system
(see Figure 2).
In our lexical alignment method, we first use
GIZA++ (Och and Ney, 2003) implementing
IBM models (Brown et al., 1993). These models
build word-based alignments from aligned sentences. Indeed, each source word has zero, one or
more translation equivalents in the target language. As these models do not provide many-tomany alignments, we also use some heuristics
(Koehn et al., 2003; Tufiş et al., 2005) to detect
phrase-based alignments such as chunks: nominal, adjectival, verbal, adverbial or prepositional
phrases.
In our experiments, we use the lemmatized,
tagged and annotated parallel corpus described in
section 2. Thus, we use lemmas and morphosyntactic properties to improve the lexical alignment. Lemmas are followed by the two first characters of morpho-syntactic tag. This operation
morphologically disambiguates the lemmas
(Tufiş et al., 2005). For example, the same
French lemma change (=exchange, modify) can
be a common noun or a verb: change_Nc vs.
change_Vm. This disambiguation procedure improves the GIZA++ system’s performance.
We realize bidirectional alignments (FR - RO
and RO - FR) with GIZA++, and we intersect
them (Koehn et al., 2003) to select common
alignments.
To improve the word alignment results, we add
an external list of cognates to the list of the translation equivalents extracted by GIZA++. This list
of cognates is built from parallel corpora by our
own method (described in the next section).
Also, to complete word alignments, we use a
French - Romanian dictionary of verbo-nominal
collocations (Todiraşcu et al., 2008). They
represent multiword expressions, composed of
words related by lexico-syntactic relations
(Todiraşcu et al., 2008). The dictionary contains
the most frequent verbo-nominal collocations
extracted from legal corpora.
To augment the recall of the lexical alignment
method, we apply a set of linguisticallymotivated heuristic rules (Tufiş et al., 2005):
a) we define some POS affinity classes (a
noun might be translated by a noun, a
verb or an adjective);
b) we align content-words such as nouns,
adjectives, verbs, and adverbs, according
to the POS affinity classes;
c) we align chunks containing translation
equivalents aligned in a previous step;
d) we align elements belonging to chunks
by linguistic heuristics. We develop a
language dependent module applying 27
morpho-syntactic contextual heuristic
rules (Navlea and Todiraşcu, 2010).
These rules are defined according to
morpho-syntactic differences between
French and Romanian.
The architecture of the lexical alignment system
is presented in the Figure 2.
Figure 2 Lexical alignment system architecture
4
Cognate Identification Module
In our hybrid cognate identification method, we
use the legal parallel corpus described in section
2. This corpus is tokenized, lemmatized, tagged,
and sentence-aligned.
Thus, we consider as cognates bilingual word
pairs respecting the linguistic conditions below:
1) their lemmas are translation equivalents
in two parallel sentences;
2) they have identical lemmas or have orthographic or phonetic similarities between lemmas;
3) they are content-words (nouns, verbs,
adverbs, etc.) having the same POS tag
or belonging to the same POS affinity
class. We filter out short words such as
prepositions and conjunctions to limit
noisy output. We also detect short cognates such as il ’he’ vs. el (personal pronoun), cas 'case' vs. caz (nouns). We
249
avoid ambiguous pairs such as lui ‘him’
(personal pronoun) (FR) vs. lui ‘s' (possessive determiner) (RO), ce 'this' (demonstrative determiner) (FR) vs. ce 'that'
(relative pronoun) (RO).
To detect orthographic and phonetic similarities
between cognates, we look at the beginning of
the words and we ignore their endings.
We classify the French - Romanian cognates detected in the studied parallel corpus (at the orthographic or phonetic level), in several categories:
1) cross-lingual invariants (numbers, certain acronyms and abbreviations, punctuation signs);
2) identical cognates (document ‘document’
vs. document);
3) similar cognates:
a) 4-grams (Simard et al., 1992);
The first 4 characters of lemmas
are identical. The length of these
lemmas is greater than or equal
to 4 (autorité vs. autoritate
'authority').
b) 3-grams; The first 3 characters
of lemmas are identical and the
length of the lemmas is greater
than or equal to 3 (acte vs. act
'paper').
c) 8-bigrams; Lemmas have a
common sequence of characters
Levels of orthographic adjustments
diacritics
double contiguous letters
consonant groups
q
intervocalic s
w
y
among the first 8 bigrams. At
least one character of each bigram is common to both words.
This condition allows the jump
of a non identical character
(souscrire vs. subscrie 'submit').
This method applies only to long
lemmas (length greater than 7).
d) 4-bigrams; Lemmas have a
common sequence of characters
among the 4 first bigrams. This
method applies for long lemmas
(length greater than 7) (homologué vs. omologat 'homologated')
but also for short lemmas (length
less than or equal to 7) (groupe
vs. grup 'group').
We iteratively extract cognates by identified categories. In addition, we use a set of orthographic
adjustments and some input data disambiguation
strategies. We compute frequency for ambiguous
candidates (the same source lemma occurs with
several target candidates) and we keep the most
frequent candidate. At each iteration, we delete
reliable considered cognates from the input data.
We start by applying a set of empirically established orthographic adjustments between French
- Romanian lemmas, such as: diacritic removal,
phonetic mappings detection, etc. (see Table 1).
French
Romanian
x
x
x
x
ph
th
dh
cch
ck
cq
ch
ch
q (final)
qu(+i) (medial)
qu(+e) (medial)
qu(+a)
que (final)
v+s+v
w
y
f [f]
t [t]
d [d]
c [k]
c [k]
c [k]
ş [∫]
c [k]
c [k]
c [k]
c [k]
c(+a) [k]
c [k]
v+z+v
v
i
Examples
FR - RO
dépôt - depozit
rapport - raport
phase - fază
méthode - metodă
adhérent - aderent
bacchante - bacantă
stockage - stocare
grecque - grec
fiche - fişă
chapitre - capitol
cinq - cinci
équilibre - echilibru
marquer - marca
qualité - calitate
pratique - practică
présent - prezent
wagon - vagon
yaourt - iaurt
Table 1 French - Romanian cognate orthographic adjustments
While French uses an etymological writing and
Romanian generally has a phonetic writing, we
identify phonetic correspondences between lemmas. Then, we make some orthographic adjustments from French to Romanian. For example,
250
Romanian cognates have one form in French, but
two various forms in Romanian (spécification
'specification' vs. specificare or specificaţie). We
recover these pairs by using regular expressions
based on specific lemma endings (ion (fr) vs.
re|ţie (ro)).
Then, we delete the reliable cognate pairs (high
precision) from the input data at the end of the
extraction step. This step helps us to disambiguate the input data. For example, the identical
cognates transport vs. transport 'transportation',
obtained in a previous extraction step and deleted
from the input data, eliminate the occurrence of
candidate transport vs. tranzit as 4-grams cognate, in a next extraction step.
We apply the same method for cognates having
POS affinity (N-V; N-ADJ). We keep only 4grams cognates, due to the significant decrease
of the precision for the other categories 3 (b, c,
d).
Deletion
Content-words /
Frequency
Precision
from the
Same POS
(%)
input data
x
100
x
x
100
cognates stockage 'stock' (FR) vs. stocare (RO)
become stocage (FR) vs. stocare (RO). In this
example, the French consonant group ck [k] become c [k] (as in Romanian). We also make adjustments in the ambiguous cases, by replacing
with both variants (ch ([∫] or [k])): fiche vs. fişă
‘sheet’; chapitre vs. capitol ‘chapter’.
We aim to improve the precision of our method.
Thus, we iteratively extract cognates by identified categories from the surest ones to less sure
candidates (see Table 2).
To decrease the noise of the cognate identification method, we apply two supplementary strategies. We filter out ambiguous cognate candidates
(autorité - autoritate|autorizare), by computing
their frequencies in the corpus. In this case, we
keep the most frequent candidate pair. This strategy is very effective to augment the precision of
the results, but it might decrease the recall in certain cases. Indeed, there are cases where French Extraction steps by category
of cognates
1 : cross lingual invariants
2 : identical cognates
3 : 4-grams (lemmas’ length
>= 4) ;
4 : 3-grams (lemmas’ length
>=3) ;
5 : 8-bigrams (long lemmas,
lemmas’ length >7)
x
x
x
99.05
x
x
x
93.13
x
95.24
x
6 : 4-bigrams (long lemmas,
lemmas’ length > 7)
x
7 : 4-bigrams (short lemmas,
lemmas’ length =< 7)
x
75
x
65.63
Table 2 Precision of cognate extraction steps
empirically establish the threshold of
0.68.
5
Evaluation and Methods’ Comparison
LCSR ( w1, w 2 ) =
We evaluated our cognate identification module
against a list of cognates initially built from the
test corpus, containing 2,034 pairs of cognates.
In addition, we also compared the results of our
method with the results provided by pure statistical methods (see Table 3). These methods are
the following:
a) thresholding the Longest Common Subsequence Ratio (LCSR) for two words of
a bilingual pair; This measure computes
the ratio between the longest common
subsequence of characters of two words
and the length of the longest word. We
length ( common _ substring ( w1, w 2 ))
max( length ( w1), length ( w 2 ))
b) thresholding DICE’s coefficient ; We
empirically establish the threshold of
0.62.
DICE ( w1, w 2 ) =
2 * number _ common _ bigrams
total _ number _ bigrams ( w1, w 2 )
c) 4-grams ; Two words are considered as
cognates if they have at least 4 characters and their first 4 characters are identical.
251
We implemented these methods using orthographically adjusted parallel corpus (see Table
1). Moreover, we evaluate 4-grams method on
the initial parallel corpus and on the orthographically adjusted parallel corpus to study the impact
of orthographic adjustments step on the quality
of the results.
These methods generally apply for words having
at least 4 letters in order to decrease the noise of
the results. Cognates are searched in aligned parallel sentences. Word characters are almost parallel (rembourser vs. rambursare 'refund').
Methods
P (%)
R (%)
F (%)
LCSR
44.13
58.95
50.47
DICE
56.47
60.91
58.61
4-grams
91.55
72.42
80.87
Our method
94.78
89.18
91.89
Table 3 Evaluation and methods’ comparison;
P=Precision; R=Recall; F=F-measure
Our method extracted 1,814 correct cognates
from 1,914 provided candidates. The method
obtains the best scores (precision=94.78% ; recall=89.18% ; f-measure=91.89%), in comparison with the other implemented methods. The 4grams method obtains a high precision (90.85%),
but a low recall (47.84%). Orthographic adjustments step improves significantly the recall of 4grams method with 24.58% (see Table 4). This
result is due to the specific properties of the law
parallel corpus. Indeed, many Romanian terms
were borrowed from French and these terms
present high orthographic similarities.
Methods
P (%)
R (%)
F (%)
4-grams Adjustments
90.85
47.84
62.68
4-grams +
Adjustments
91.55
72.42
80.87
Table 4 Evaluation of the 4-grams method before and
after orthographic adjustments step
However, our method extracts some ambiguous
candidates such as numéro ‘number’ - nume
‘name’, compléter ‘complete’ - compune ‘compose’. Some of these errors were avoided by
keeping the most frequent candidate in the studied corpus. So, the remaining errors mainly
concern hapax candidates.
Also, some cognates were not extracted: heure oră ‘hour’, semaine - săptămână ‘week’, lieu loc ‘place’. These errors concern cognates sharing very few orthographic similarities.
The lowest scores are obtained by the LCSR method (f-measure=50.47%), followed by the
DICE’s coefficient (f-measure=58.61%). These
general methods provide a high noise due to the
important orthographic similarities between the
words having different meanings. Their results
might be improved by combining statistical techniques with linguistic information such as POS
affinity or by combining several association
scores.
As we mentioned, the output of the cognate identification module is exploited by a French - Romanian lexical alignment system (based on GIZA++) described in section 3. We compared the
set of cognates provided by GIZA++ with our
results to study their impact on cognate alignment. GIZA++ extracted 1,532 cognates
representing a recall of 75.32% (see Table 5).
Our cognate identification module significantly
improved the recall with 13.86%.
Number of
extracted
cognates
1,532
GIZA++
Our
1,814
method
Systems
Number
of total
cognates
Recall
(%)
75.32
2,034
89.18
Table 5 Improvement of our method’s recall
6
Conclusions and Further Work
We present a French - Romanian cognate identification module required by a lexical alignment
system. Our method combines statistical techniques and linguistic filters to extract cognates
from lemmatized, tagged and sentence-aligned
parallel corpus. The use of the linguistic information and the orthographic adjustments significantly improves the results compared with pure
statistical methods. However, these results are
dependent of the studied languages, of the corpus
domain and of the data volume. We need more
experiments using other corpora from other domains to be able to generalize. Our system
should be improved to detect false friends by
using external resources.
252
tical Machine Translation Systems, in Proceedings
of the Workshop on Exploitation of Multilingual
Resources and Tools for Central and (South) Eastern European Languages, 7th International Conference on Language Resources and Evaluation,
Malta, Valletta, May 2010, pp. 41-48.
Cognate identification module will be integrated
in a French - Romanian lexical alignment system. This system is part of a larger project aiming to develop a factored phrase-based statistical
machine translation system for French and Romanian.
References
George W. Adamson and Jillian Boreham. 1974. The
use of an association measure based on character
structure to identify semantically related pairs of
words and document titles, Information Storage
and Retrieval, 10(7-8):253-260.
Chris Brew and David McKelvie. 1996. Word-pair
ex-traction for lexicography, in Proceedings of International Conference on New Methods in Natural
Language Processing, Bilkent, Turkey, 45-55.
Peter F. Brown, Vincent J. Della Pietra, Stephen A.
Della Pietra, and Robert L. Mercer. 1993. The mathematics of statistical machine translation: Parameter estimation, Computational Linguistics,
19(2):263-312.
Nancy Ide and Jean Véronis. 1994. Multext (multilingual tools and corpora), in Proceedings of the 15th
International Conference on Computational Linguistics, CoLing 1994, Kyoto, pp. 90-96.
Diana Inkpen, Oana Frunză, and Grzegorz Kondrak.
2005. Automatic Identification of Cognates and
False Friends in French and English, RANLP2005, Bulgaria, Sept. 2005, p. 251-257.
Radu Ion. 2007. Metode de dezambiguizare semantică
automată. Aplicaţii pentru limbile engleză şi
română, Ph.D. Thesis, Romanian Academy, Bucharest, May 2007, 148 pp.
Philipp Koehn, Franz J. Och, and Daniel Marcu.
2003. Statistical Phrase-Based Translation, in Proceedings of Human Language Technology Conference of the North American Chapter of the Association of Computational Linguistics, HLT-NAACL
2003, Edmonton, May-June 2003, pp. 48-54.
Grzegorz Kondrak. 2009. Identification of Cognates
and Recurrent Sound Correspondences in Word
Lists, in Traitement Automatique des Langues
(TAL), 50(2) :201-235.
Olivier Kraif. 1999. Identification des cognats et alignement bi-textuel : une étude empirique, dans
Actes de la 6ème conférence annuelle sur le Traitement Automatique des Langues Naturelles, TALN
99, Cargèse, 12-17 juillet 1999, 205-214.
Dan I. Melamed. 1999. Bitext Maps and Alignment
via Pattern Recognition, in Computational Linguistics, 25(1):107-130.
Michael P. Oakes. 2000. Computer Estimation of Vocabulary in Protolanguage from Word Lists in Four
Daughter Languages, in Journal of Quantitative
Linguistics, 7(3):233-243.
Franz J. Och and Hermann Ney. 2003. A Systematic
Comparison of Various Statistical Alignment
Models, in Computational Linguistics, 29(1):1951.
Michel Simard, George Foster, and Pierre Isabelle.
1992. Using cognates to align sentences, in Proceedings of the Fourth International Conference on
Theoretical and Methodological Issues in Machine
Translation, Montréal, pp. 67-81.
Amalia Todiraşcu, Ulrich Heid, Dan Ştefănescu, Dan
Tufiş, Christopher Gledhill, Marion Weller, and
François Rousselot. 2008. Vers un dictionnaire de
collocations multilingue, in Cahiers de Linguistique, 33(1) :161-186, Louvain, août 2008.
Amalia Todiraşcu, Radu Ion, Mirabela Navlea, and
Laurence Longo. 2011. French text preprocessing
with TTL, in Proceedings of the Romanian Academy, Series A, Volume 12, Number 2/2011, pp.
151-158, Bucharest, Romania, June 2011, Romanian Academy Publishing House. ISSN 1454-9069.
Dan Tufiş and Ana Maria Barbu. 1997. A Reversible
and Reusable Morpho-Lexical Description of Romanian, in Dan Tufiş and Poul Andersen (eds.),
Recent Advances in Romanian Language Technology, pp. 83-93, Editura Academiei Române,
Bucureşti, 1997. ISBN 973-27-0626-0.
Dan Tufiş, Radu Ion, Alexandru Ceauşu, and Dan
Ştefănescu. 2005. Combined Aligners, in Proceedings of the Workshop on Building and Using Parallel Texts: Data-Driven Machine Translation and
Beyond, pp. 107-110, Ann Arbor, USA, Association for Computational Linguistics. ISBN 978-973703-208-9.
Cristina Vertan and Monica Gavrilă. 2010. Multilingual applications for rich morphology language
pairs, a case study on German Romanian, in Dan
Tufiş and Corina Forăscu (eds.): Multilinguality
and Interoperability in Language Processing with
Emphasis on Romanian, Romanian Academy Publishing House, Bucharest, pp. 448-460, ISBN 978973-27-1972-5.
Robert A. Wagner and Michael J. Fischer. 1974. The
String-to-String Correction Problem, Journal of the
ACM, 21(1):168-173.
Mirabela Navlea and Amalia Todiraşcu. 2010. Linguistic Resources for Factored Phrase-Based Statis-
253